Regularized discriminative broad learning system for image classification

Because of its simple network structure and efficient learning mode, the Broad Learning System (BLS) has achieved impressive performance in image classification tasks. Nevertheless, two deficiencies still exist which have severely limited its learning ability. First, the strict binary labeling strategy used in BLS-based models restricts the model’s flexibility. Second, the final broad features are inevitably redundant, which can cause useless features to be learned and reduce the recognition accuracy. In this paper, we propose three discriminative BLS-based models to address these mentioned problems. Specifically, we first integrate the ɛ-dragging technique into the framework of standard BLS to relax the regression targets and propose the ℓ2-norm based discriminative BLS (L2DBLS) model. Secondly, to avoid the negative effects of redundant features in L2DBLS, we utilize the ℓ2,1 regularizer to replace the Frobenius norm for feature selection. Furthermore, we propose to constrain the projection matrix of BLS by ℓ2 and ℓ2,1 regularization simultaneously. As a result, the obtained output weights can be more compact and smooth for recognition. Efficient iterative methods based on the alternating direction method of multipliers are derived to optimize the proposed models. Finally, various experiments on image databases are intended to demonstrate the outstanding recognition capability of our proposed models in comparison with other state-of-the-art classifiers.